How to set max-out groundtruth in training?
Closed this issue · 1 comments
chenpengf0223 commented
@sfzhang15
Thanks for sharing your great work.
I have some confusion about the 'max-out' background label, for example I use Nm=3, the groudtruth of the label should be "0, 0, 0, 1"for positive example, and "1, 1, 1, 0" for negative example?
If there is the process in the code?
sfzhang15 commented
@chenpengf0223
The max-out operation first outputs four channels, then uses max eltwise opeation to output two channels. Thus, the ground-truth label does not need to change, which is the same as the original set. Please download our trained model and prototxt for more details.